Emergence and Emergent Capabilities

Emergence is when a system exhibits properties its parts don't have — "more is different." In AI, emergent capabilities are skills that appear suddenly as models scale, without being explicitly trained in.

#The surprise

Below some size, a model can't do a task at all; past a threshold, it suddenly can — arithmetic, multi-step reasoning, translation, in-context learning. No one programmed these; they fell out of scale.

small model:  can't do X (≈ chance)
              ... scale up ...
large model:  can do X    ← emerged, unplanned

#Why it's profound (and unsettling)

• We can't fully predict what a bigger model will be able to do — capability is partly discovered after training.
• That cuts both ways for safety: dangerous capabilities can emerge unexpectedly too.
• It fuels debate about AGI: is "understanding" just another capability waiting to emerge at scale, or a category difference (The Chinese Room)?

Caveat: some "emergence" is partly an artifact of how we measure it (sharp metrics make smooth gains look like jumps) — a live scientific debate, not settled magic.

Related: How Transformers Work · The Sharp Left Turn · Strong vs Weak AI (ANI, AGI, ASI)